AI model training that fits your content
Training AI models doesn’t have to be a complex process locked behind layers of code and data science jargon. With DeepVA, you can train your own AI model — faces, voices, landmarks — using the media you already have.
custom face training
Train faces you know. Recognize the people who matter.
With DeepVA’s Face Training model, identifying familiar faces across your video archives becomes effortless. Just upload a few clips or images — the system handles the rest. It crops, detects, and starts training your custom model using DeepVA’s Deep Model Customizer.
custom speaker training
Give voice to your content: Speaker Training
custom landmark training
Places have voices too: Landmark Training
Some stories are tied to where they happened. Landmark Training helps you train AI models that spot specific locations — like city halls, cultural landmarks, or branded spaces — straight from your own footage or imagery.
Dataset evaluation
AI model evaluation
With DeepVA you can automatically evaluate the quality of your training data for face and landmark recognition. The system checks factors like image resolution, relative size, and positioning of relevant features—providing clear feedback to help you improve your dataset before training. This ensures more accurate and reliable AI model results, without requiring any machine learning expertise.
Custom AI model training module is part of our Deep Media Customizer application. Check it out now:
The benefits of DeepVA custom AI model training
Adapt the AI to your needs
Tailor recognition models for faces, voices, and landmarks to reflect company-specific, regional, or contextual needs.
No-code training interface for fast deployment
Quickly build custom AI models through a no-code interface – even with very limited training data, thanks to few-shot learning.
Seamless integration with media workflows
Trained models can be deployed immediately across the DeepVA ecosystem or integrated with third-party systems via API for consistent cross-platform recognition.
Practical Applications
Custom AI model training
Real-world success stories
rbb
rbb’s collaboration with DeepVA transformed regional broadcasting and archiving through efficient AI utilization.
Heilbronn
DeepVA helps the Heilbronn City Archive to automatically index and digitize a large amount of images.
Bayerischer Rundfunk
Creating and managing training data for face recognition using AI requires a lot of time and resources.
frequently asked questions
Have a question? We’ve got answers
What is AI Model Training?
AI model training is the process of teaching an artificial intelligence system how to recognize specific patterns — like faces, voices, or landmarks — using examples that you provide. You upload your own media (images, video clips, audio), label what’s relevant, and the system learns to identify those patterns in future content.
But here’s something important: when you train an AI model with DeepVA, your data stays yours. We don’t store, reuse, or learn from your training data — ever. The models you train are isolated to your organization, and all data handling is fully GDPR-compliant. No hidden training on your content, no surprise usage, and no shared datasets behind the scenes.
It’s not just smart — it’s secure. You’re in full control of your training process, your media, and your results. That means you get a custom AI model that truly fits your world — without compromising privacy or ownership.
How many identities can I train in a single model?
You can train up to several thousand individual identities, depending on the dataset and desired accuracy. The model scales flexibly.
How much data is required to train a person’s face?
As little as 5–10 high-quality images per person can yield strong results. More data improves accuracy and confidence scores.
What kinds of landmarks can be trained and how much training data is needed?
Anything visually consistent and identifiable — including buildings, monuments, landscapes, and even brand-specific environments or interior spaces.
Typically, 20–50 high-quality images per landmark provide a solid training base. More images improve the precision and confidence score of detection.
How many speakers can I train in a single model?
The system supports hundreds of trained speaker profiles. Larger-scale deployments can be customized based on your infrastructure and performance needs.
How much audio is needed to train a speaker? How many speakers can I train in a single model?
Around 2–5 minutes of clear speech per person is sufficient for reliable training. More data improves confidence and accuracy.
Can I combine Speaker Training with transcription tools?
Yes. The model integrates seamlessly with transcription workflows, attributing each segment to the correct speaker for cleaner, more structured output — you can read more about audio transcription here.
Can I update or retrain an existing model?
Absolutely. You can add new faces, remove outdated ones, or retrain models as your needs evolve — without starting from scratch.
Is your service GDPR compliant?
Yes, DeepVA is fully GDPR compliant. We take data protection and privacy seriously and ensure that all personal data is processed in accordance with GDPR regulations.
How is my data handled? Does the AI learn from my data?
You have full control over your data on our AI platform, ensuring it remains secure and compliant. By default, we do not use your data to train our models, keeping it proprietary. However, you have the option to train models using your data, and in that case, it will remain exclusive to your organization.
What type of data do you store?
By default, we do not process your data beyond what is required to provide our services. If additional processing is necessary, it will only occur as outlined in your instructions or where legally required. For example, data may be transferred or processed as needed to fulfill service requirements, always in alignment with our agreements.
To learn more about how we process data and the safeguards in place, please refer to our Data Processing Agreement.
